Crash less. The i64 restriction in BinomialCoefficient caused some problems

with code that was expecting different bit widths for different values.

Make getTruncateOrZeroExtend a method on ScalarEvolution, and use it.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52248 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Analysis/ScalarEvolution.h

@@ -253,6 +253,11 @@ namespace llvm {
     SCEVHandle getMinusSCEV(const SCEVHandle &LHS,
                             const SCEVHandle &RHS);
 
+    /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion
+    /// of the input value to the specified type.  If the type must be
+    /// extended, it is zero extended.
+    SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty);
+
     /// getIntegerSCEV - Given an integer or FP type, create a constant for the
     /// specified signed integer value and return a SCEV for the constant.
     SCEVHandle getIntegerSCEV(int Val, const Type *Ty);

lib/Analysis/ScalarEvolution.cpp

@@ -499,21 +499,6 @@ SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
   return getUnknown(C);
 }
 
-/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
-/// input value to the specified type.  If the type must be extended, it is zero
-/// extended.
-static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty,
-                                          ScalarEvolution &SE) {
-  const Type *SrcTy = V->getType();
-  assert(SrcTy->isInteger() && Ty->isInteger() &&
-         "Cannot truncate or zero extend with non-integer arguments!");
-  if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
-    return V;  // No conversion
-  if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
-    return SE.getTruncateExpr(V, Ty);
-  return SE.getZeroExtendExpr(V, Ty);
-}
-
 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
 ///
 SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
@@ -585,7 +570,7 @@ static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
 #endif
 
   const IntegerType *DividendTy = IntegerType::get(DividendBits);
-  const SCEVHandle ExIt = SE.getZeroExtendExpr(It, DividendTy);
+  const SCEVHandle ExIt = SE.getTruncateOrZeroExtend(It, DividendTy);
 
   // The final number of bits we need to perform the division is the maximum of
   // dividend and divisor bitwidths.
@@ -607,7 +592,12 @@ static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
       Dividend *= N-(K-1);
     if (DividendTy != DivisionTy)
       Dividend = Dividend.zext(DivisionTy->getBitWidth());
-    return SE.getConstant(Dividend.udiv(Divisor).trunc(It->getBitWidth()));
+
+    APInt Result = Dividend.udiv(Divisor);
+    if (Result.getBitWidth() != It->getBitWidth())
+      Result = Result.trunc(It->getBitWidth());
+
+    return SE.getConstant(Result);
   }
   
   SCEVHandle Dividend = ExIt;
@@ -615,11 +605,12 @@ static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
     Dividend =
       SE.getMulExpr(Dividend,
                     SE.getMinusSCEV(ExIt, SE.getIntegerSCEV(i, DividendTy)));
-  if (DividendTy != DivisionTy)
+
-    Dividend = SE.getZeroExtendExpr(Dividend, DivisionTy);
+  return SE.getTruncateOrZeroExtend(
-  return
+             SE.getUDivExpr(
-    SE.getTruncateExpr(SE.getUDivExpr(Dividend, SE.getConstant(Divisor)),
+                 SE.getTruncateOrZeroExtend(Dividend, DivisionTy),
-                       It->getType());
+                 SE.getConstant(Divisor)
+             ), It->getType());
 }
 
 /// evaluateAtIteration - Return the value of this chain of recurrences at
@@ -703,6 +694,21 @@ SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op, const Type *
   return Result;
 }
 
+/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion
+/// of the input value to the specified type.  If the type must be
+/// extended, it is zero extended.
+SCEVHandle ScalarEvolution::getTruncateOrZeroExtend(const SCEVHandle &V,
+                                                    const Type *Ty) {
+  const Type *SrcTy = V->getType();
+  assert(SrcTy->isInteger() && Ty->isInteger() &&
+         "Cannot truncate or zero extend with non-integer arguments!");
+  if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
+    return V;  // No conversion
+  if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
+    return getTruncateExpr(V, Ty);
+  return getZeroExtendExpr(V, Ty);
+}
+
 // get - Get a canonical add expression, or something simpler if possible.
 SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
   assert(!Ops.empty() && "Cannot get empty add!");
@@ -2432,8 +2438,8 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) {
       // loop iterates.  Compute this now.
       SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
       if (IterationCount == UnknownValue) return UnknownValue;
-      IterationCount = getTruncateOrZeroExtend(IterationCount,
+      IterationCount = SE.getTruncateOrZeroExtend(IterationCount,
-                                               AddRec->getType(), SE);
+                                                  AddRec->getType());
 
       // If the value is affine, simplify the expression evaluation to just
       // Start + Step*IterationCount.

test/Analysis/ScalarEvolution/2008-06-12-BinomialInt64.ll

@@ -0,0 +1,43 @@
+; RUN: llvm-as < %s | opt -analyze -scalar-evolution 2>/dev/null
+; PR2433
+
+define i32 @main1(i32 %argc, i8** %argv) nounwind  {
+entry:
+	br i1 false, label %bb10, label %bb23
+
+bb10:		; preds = %bb10, %entry
+	%accum.03 = phi i64 [ %tmp14, %bb10 ], [ 0, %entry ]		; <i64> [#uses=1]
+	%i.02 = phi i32 [ %tmp16, %bb10 ], [ 0, %entry ]		; <i32> [#uses=1]
+	%d.1.01 = phi i64 [ %tmp5.i, %bb10 ], [ 0, %entry ]		; <i64> [#uses=1]
+	%tmp5.i = add i64 %d.1.01, 1		; <i64> [#uses=2]
+	%tmp14 = add i64 %accum.03, %tmp5.i		; <i64> [#uses=2]
+	%tmp16 = add i32 %i.02, 1		; <i32> [#uses=2]
+	%tmp20 = icmp slt i32 %tmp16, 0		; <i1> [#uses=1]
+	br i1 %tmp20, label %bb10, label %bb23
+
+bb23:		; preds = %bb10, %entry
+	%accum.0.lcssa = phi i64 [ 0, %entry ], [ %tmp14, %bb10 ]		; <i64> [#uses=0]
+	ret i32 0
+}
+
+define i32 @main2(i32 %argc, i8** %argv) {
+entry:
+	%tmp8 = tail call i32 @atoi( i8* null ) nounwind readonly 		; <i32> [#uses=1]
+	br i1 false, label %bb9, label %bb21
+
+bb9:		; preds = %bb9, %entry
+	%accum.03 = phi i64 [ %tmp12, %bb9 ], [ 0, %entry ]		; <i64> [#uses=1]
+	%i.02 = phi i32 [ %tmp14, %bb9 ], [ 0, %entry ]		; <i32> [#uses=1]
+	%d.1.01 = phi i64 [ %tmp4.i, %bb9 ], [ 0, %entry ]		; <i64> [#uses=1]
+	%tmp4.i = add i64 %d.1.01, 1		; <i64> [#uses=2]
+	%tmp12 = add i64 %accum.03, %tmp4.i		; <i64> [#uses=2]
+	%tmp14 = add i32 %i.02, 1		; <i32> [#uses=2]
+	%tmp18 = icmp slt i32 %tmp14, %tmp8		; <i1> [#uses=1]
+	br i1 %tmp18, label %bb9, label %bb21
+
+bb21:		; preds = %bb9, %entry
+	%accum.0.lcssa = phi i64 [ 0, %entry ], [ %tmp12, %bb9 ]		; <i64> [#uses=0]
+	ret i32 0
+}
+
+declare i32 @atoi(i8*) nounwind readonly